Awesome

Steam StyleGAN

The goal of this Google Colab notebook is to capture the distribution of Steam banners and sample with a StyleGAN.

Usage

• Acquire the data, e.g. as a snapshot called 128x128.zip in another of my repositories,
• Follow the instructions to edit train.py in the official StyleGAN Github repository,
• Run StyleGAN.ipynb to train a StyleGAN.
• To resume training from a checkpoint, you will have to edit training/training_loop.py.

NB: You might have to edit metrics/frechet_inception_distance.py to retrieve the network inception_v3_features.pkl locally if it cannot be downloaded from Google Colab.

Results

The dataset consists of 31,723 Steam banners with RGB channels and resized from 460x215 to 128x128 resolution. Pre-processed data, as .tfrecords files, can be downloaded from Google Drive.

A StyleGAN model was trained on 3,524,000 images, with a decreasing mini-batch size, which is about 111 epochs. A checkpoint of the network can be downloaded from Google Drive.

Caveat: training was manually stopped after roughly 1 day, using 1 Tesla K80 GPU in the cloud. Based on the expected training times for 1024x1024, 512x512 and 256x256 images, 9 days of computation time might be required to get the best results for 128x128 images.

Generated Steam banners

Results obtained with different numbers of images seen during training are shown on the Wiki.

A grid of generated Steam banners after 3,524 kimg:

Real Steam banners

A grid of real Steam banners:

References

• StyleGAN2:
  • StyleGAN2
  • Steam-StyleGAN2
• StyleGAN:
  • StyleGAN1
• DCGAN:
  • Steam-DCGAN